kubelet container status calculation doesn't handle suddenly missing data properly

Author: deads2k

pkg/kubelet/kubelet_pods.go

@@ -1616,7 +1616,7 @@ func (kl *Kubelet) convertStatusToAPIStatus(pod *v1.Pod, podStatus *kubecontaine
 // convertToAPIContainerStatuses converts the given internal container
 // statuses into API container statuses.
 func (kl *Kubelet) convertToAPIContainerStatuses(pod *v1.Pod, podStatus *kubecontainer.PodStatus, previousStatus []v1.ContainerStatus, containers []v1.Container, hasInitContainers, isInitContainer bool) []v1.ContainerStatus {
-	convertContainerStatus := func(cs *kubecontainer.Status) *v1.ContainerStatus {
+	convertContainerStatus := func(cs *kubecontainer.Status, oldStatus *v1.ContainerStatus) *v1.ContainerStatus {
 		cid := cs.ID.String()
 		status := &v1.ContainerStatus{
 			Name:         cs.Name,
@@ -1625,17 +1625,17 @@ func (kl *Kubelet) convertToAPIContainerStatuses(pod *v1.Pod, podStatus *kubecon
 			ImageID:      cs.ImageID,
 			ContainerID:  cid,
 		}
-		switch cs.State {
-		case kubecontainer.ContainerStateRunning:
+		switch {
+		case cs.State == kubecontainer.ContainerStateRunning:
 			status.State.Running = &v1.ContainerStateRunning{StartedAt: metav1.NewTime(cs.StartedAt)}
-		case kubecontainer.ContainerStateCreated:
+		case cs.State == kubecontainer.ContainerStateCreated:
 			// Treat containers in the "created" state as if they are exited.
 			// The pod workers are supposed start all containers it creates in
 			// one sync (syncPod) iteration. There should not be any normal
 			// "created" containers when the pod worker generates the status at
 			// the beginning of a sync iteration.
 			fallthrough
-		case kubecontainer.ContainerStateExited:
+		case cs.State == kubecontainer.ContainerStateExited:
 			status.State.Terminated = &v1.ContainerStateTerminated{
 				ExitCode:    int32(cs.ExitCode),
 				Reason:      cs.Reason,
@@ -1644,6 +1644,24 @@ func (kl *Kubelet) convertToAPIContainerStatuses(pod *v1.Pod, podStatus *kubecon
 				FinishedAt:  metav1.NewTime(cs.FinishedAt),
 				ContainerID: cid,
 			}
+
+		case cs.State == kubecontainer.ContainerStateUnknown &&
+			oldStatus != nil && // we have an old status
+			oldStatus.State.Running != nil: // our previous status was running
+			// if this happens, then we know that this container was previously running and isn't anymore (assuming the CRI isn't failing to return running containers).
+			// you can imagine this happening in cases where a container failed and the kubelet didn't ask about it in time to see the result.
+			// in this case, the container should not to into waiting state immediately because that can make cases like runonce pods actually run
+			// twice. "container never ran" is different than "container ran and failed".  This is handled differently in the kubelet
+			// and it is handled differently in higher order logic like crashloop detection and handling
+			status.State.Terminated = &v1.ContainerStateTerminated{
+				Reason:   "ContainerStatusUnknown",
+				Message:  "The container could not be located when the pod was terminated",
+				ExitCode: 137, // this code indicates an error
+			}
+			// the restart count normally comes from the CRI (see near the top of this method), but since this is being added explicitly
+			// for the case where the CRI did not return a status, we need to manually increment the restart count to be accurate.
+			status.RestartCount = oldStatus.RestartCount + 1
+
 		default:
 			// this collapses any unknown state to container waiting.  If any container is waiting, then the pod status moves to pending even if it is running.
 			// if I'm reading this correctly, then any failure to read status on any container results in the entire pod going pending even if the containers
@@ -1767,7 +1785,11 @@ func (kl *Kubelet) convertToAPIContainerStatuses(pod *v1.Pod, podStatus *kubecon
 		if containerSeen[cName] >= 2 {
 			continue
 		}
-		status := convertContainerStatus(cStatus)
+		var oldStatusPtr *v1.ContainerStatus
+		if oldStatus, ok := oldStatuses[cName]; ok {
+			oldStatusPtr = &oldStatus
+		}
+		status := convertContainerStatus(cStatus, oldStatusPtr)
 		if containerSeen[cName] == 0 {
 			statuses[cName] = status
 		} else {

pkg/kubelet/kubelet_test.go

@@ -1607,6 +1607,38 @@ func TestGenerateAPIPodStatusWithReasonCache(t *testing.T) {
 				}},
 			},
 		},
+		// For Unknown Container Status:
+		// * In certain situations a container can be running and fail to retrieve the status which results in
+		// * a transition to the Unknown state. Prior to this fix, a container would make an invalid transition
+		// * from Running->Waiting. This test validates the correct behavior of transitioning from Running->Terminated.
+		{
+			containers: []v1.Container{{Name: "unknown"}},
+			statuses: []*kubecontainer.Status{
+				{
+					Name:  "unknown",
+					State: kubecontainer.ContainerStateUnknown,
+				},
+				{
+					Name:  "unknown",
+					State: kubecontainer.ContainerStateRunning,
+				},
+			},
+			reasons: map[string]error{},
+			oldStatuses: []v1.ContainerStatus{{
+				Name:  "unknown",
+				State: v1.ContainerState{Running: &v1.ContainerStateRunning{}},
+			}},
+			expectedState: map[string]v1.ContainerState{
+				"unknown": {Terminated: &v1.ContainerStateTerminated{
+					ExitCode: 137,
+					Message:  "The container could not be located when the pod was terminated",
+					Reason:   "ContainerStatusUnknown",
+				}},
+			},
+			expectedLastTerminationState: map[string]v1.ContainerState{
+				"unknown": {Running: &v1.ContainerStateRunning{}},
+			},
+		},
 	}
 
 	for i, test := range tests {